rvc-genshin-impact / README.md
ArkanDash's picture
feat: added kazuha-jp
3e2cf66
|
raw
history blame
2.16 kB
metadata
license: mit
language:
  - ja
metrics:
  - accuracy
pipeline_tag: audio-to-audio
tags:
  - rvc

RVC Genshin Impact Japanese Voice Model

model-cover.png

About Retrieval based Voice Conversion (RVC)

Learn more about Retrieval based Voice Conversion in this link below:
RVC WebUI

How to use?

Download the prezipped model and put to your RVC Project.

Model test: Google Colab / RVC Models New (Which is basically the same but hosted on spaces)

INFO

Model Created by ArkanDash
The voice that was used in this model belongs to Hoyoverse.

The voice I make to make this model was ripped from the game (3.7).
Total Models: 33 Models (19 V1 Models & 14 V2 Models)

V1 Model

This was trained on Original RVC.
Pitch Extract using Harvest.
This model was trained with 100 epochs, 10 batch sizes, and a 40K sample rate (some models had a 48k sample rate).
Every V1 model was trained more or less around 30 minutes of character voice.

I may exclude some models to higher epochs due to the low duration of the character's voice.

  • Klee 150 Epochs
  • Fischl 150 Epochs

(New) V2 Model

This was trained on Mangio-Fork RVC.
Pitch Extract using Crepe.
This model was trained with 100 epochs, 8 batch sizes, and a 48K sample rate. (some models had a 40k sample rate).
Every V2 model was trained more or less around 60 minutes of character voice.

Other request:

  • Greater Lord Rukkhadevata: 750 Epochs, 16 Batch size, 48k Sample rate. (10 minutes dataset)
  • Charlotte: 400 Epochs, 16 Batch size, 48k Sample rate. (18 minutes dataset)

Note:

  • For faruzan, somehow the index file is smaller, But it output a log when training here:
    Converged (lack of improvement in inertia) at step 1152/48215
    I might retrain faruzan soon.